Construction Of CT Differential Model Between Renal Adipose-deficient Angiomyolipoma And Homogeneous Renal Cell Carcinoma Based On Fast Regional Convolution Neural Network

Objective:Renal angiomyolipoma is the most common benign renal lesion,accounting for about1.3% of the total number of renal tumors.It is common in women,of which about 80% are isolated,often small and single,with slow development.Rupture and bleeding is the mainway to endanger human health.Due to the presence of fat in most renal angiomyolipomas,it is easy to distinguish them from renal malignant tumors in imaging.However,renal angiomyolipomas lacking fat components account for about 20% of the total.It is still a difficult diagnostic problem to distinguish this type of renal space occupying lesions from homogeneous renal malignant tumors.The treatment of renal angiomyolipoma is quite different from that of renal cell carcinoma.Accurate preoperative diagnosis can reduce or even avoid unnecessary renal surgery to preserve patients’ renal function,reduce patients’ pain and reduce the waste of medical resources.The aim of our study was to explore the intelligent recognition and differentiation of renal fat-poor angiomyolipoma and homogeneous renal cell carcinoma based on fast regional convolution neural network on enhanced CT images.Methods:The contrast-enhanced CT images of 50 patients with renal fat-deficient angiomyolipoma and 150 patients with homogeneous renal cell carcinoma treated in the affiliated Hospital of Qingdao University from 2011.1 to 2021.1 were collected retrospectively.4024 images of cortical phase,parenchymal phase and excretory phase were obtained,of which 325 images were screened and 885 images of homogeneous renal cell carcinoma were selected.The image data is divided into training set and test set at the proportion of 4:1.A fast region convolution neural network framework is constructed,which is composed of feature extraction network Resnet101,region generation network RPN and ROI pooling layer.Data from the training set were introduced into a model constructed by a fast regional convolution neural network,which was trained to identify and distinguish CT images of patients with renal adipose-deficient angiomyolipoma and homogeneous renal cell carcinoma.The prediction ability of the model is evaluated by test set data,and the evaluation indicators include loss function value,accuracy,recall rate,average accuracy,average accuracy of multiple verification tests and so on.Results:The results of the fitted model performed well in the test set.The recall rate of renal fat-poor angiomyolipoma was 0.975,and the AP value was 0.958.The recall rate of homogeneous renal cell carcinoma was 0.963,and the AP value was 0.972.The m AP of the model to distinguish the above two kinds of renal tumors was 0.915,and the accuracy was94%.It takes 0.2 seconds for the model to diagnose a single image.Conclusion:The differential model based on fast regional convolution neural network can accurately distinguish renal fat-poor angiomyolipoma from homogeneous renal cell carcinoma in CT images under certain conditions.